DeepSeek shock. In many ways, it's likely to accelerate the competition of LLM & multimodal foundation models.
Dataset collection and processing areas for training LLM foundation models are largely open.
This area has a lot of changes, so the open-source power works well.
Llama is giving a lot of hints in learning the base model, and it seems that DeepSeek gives around 50% hints of reaching SOTA (the highest level of technology).
It is significant in that it has shown the "possibility" that fast followings are possible.
The fine tuning stage showed the possibility of automation as the method disclosed by DeepSeek-R1-Zero developed.
It is already seen that it is trying to recreate it centered on the open-source community.
.
DeepSeek is never a low-cost achievement. I'm sure of this.
It is an achievement achieved by DeepSeek at a low cost, but fast follow-up is impossible if the center of gravity is focused.
It's like the argument that "a country can grow without protectionism."
If the message of achievements achieved at low cost is solidified, it will play a role in blocking latecomers.
Just by looking at DeepSeek's global service, you can see that the size of their GPUs is frighteningly vast.
.
When doing important work, capital was not the first priority.
Always the most important thing was "people" who could achieve their goals.
'U.S stocks [2025] ISSUE arrangemet' 카테고리의 다른 글
ChatGPT is really good (7) | 2025.01.30 |
---|---|
새벽에 있을 FOMC 관전포인트 (10) | 2025.01.30 |
DeepSeek's low-cost learning is not an important factor. (8) | 2025.01.29 |
Netflix is said to have exceeded 300 million (6) | 2025.01.29 |
Is DeepSeek-R1 really a cost-effective model? (5) | 2025.01.29 |